At the SC26 supercomputing conference in November 2026, researchers from Lawrence Livermore National Laboratory (LLNL), Columbia University, and Advanced Micro Devices (AMD) showcased ElMerFold, an exascale protein structure distillation workflow running on the El Capitan supercomputer. Selected as a 2026 ACM Gordon Bell Prize finalist, the system generated high-accuracy 3D structural predictions for more than 41 million monomer proteins in roughly 11 hours, providing the foundational synthetic training dataset needed to develop OpenFold3 as an open-source alternative to proprietary biomolecular foundation models.
What Happened: ElMerFold at SC26 and the Gordon Bell Selection
On November 15, 2026, Lawrence Livermore National Laboratory and partners announced the operational results of ElMerFold on the supercomputer El Capitan. The technical paper, ‘Exascale Distillation Workflows for Protein Structure Prediction on El Capitan’, was selected as an ACM Gordon Bell Prize finalist. The team—led by principal investigator Brian C. Van Essen and technical lead Nikoli Dryden of LLNL, alongside Mohammed AlQuraishi of Columbia University and Nicholas Malaya of AMD—presents Tuesday, November 17, 2026, at SC26 in Salt Lake City.
ElMerFold executed the largest protein structure prediction run reported to date. Scaling across up to 11,200 nodes and 44,800 AMD Instinct MI300A Accelerated Processing Units (APUs) on El Capitan, the workflow folded roughly 41 million monomer protein sequences from the MGnify database. Production runs sustained 2,400 structures per second, while benchmark tests peaked at 6,718 structures per second. This yielded a 17.2-fold speedup over OpenFold2 baselines on prior-generation GPUs and up to 45.4-fold over earlier distributed workflows, compressing six to eight days of supercomputer time into roughly 11 hours.
‘Doing this kind of AI inference at scale and being able to take a trained model and generate lots of data for distillation—or to explore the design space for bioresilience—is absolutely critical as a capability,’ said Brian Van Essen, LLNL senior computer scientist and project principal investigator. ‘And to be able to do it on El Capitan is game-changing, because we can reach scales that you just cannot get anywhere else.’
Key Technical Architecture and I/O Optimizations
ElMerFold was tailored to exploit El Capitan’s node architecture, currently ranked No. 1 on the TOP500 list of supercomputers. Rather than linking host CPUs across PCIe buses to separate GPU accelerators, El Capitan’s compute blades feature AMD Instinct MI300A APUs. Each APU integrates 24 Zen 4 CPU cores, CDNA 3 graphics compute units, and 128 gigabytes of shared HBM3 memory on a single silicon package. This unified memory model eliminates host-to-device PCIe copy overhead, removing a primary bottleneck in memory-bound structural inference.
Software optimizations spanned the application stack. The team developed the LBANNv2 backend for PyTorch, optimizing tensor communication and kernel execution for the MI300A. Spatial memory optimizations used the DaCe (Data-Centric Parallel Programming) framework alongside customized Triton kernels. Dynamic task scheduling and node-level fault resilience were managed by LLNL’s open-source Flux workload manager.
To prevent I/O contention during 41-million-structure runs, ElMerFold routed intermediate structural data through HPE Cray near-node local NVMe storage tiers (‘Rabbit’ modules) before asynchronous transfer to the Lustre file system, reflecting trends in high-throughput supercomputing storage architectures unveiled at SC26. The multi-stage pipeline—including sequence alignment filtering and molecular dynamics relaxation—was packaged via containerized environments to ensure reproducibility across distributed nodes, aligning with considerations explored in CASRAI’s analysis of Docker versus virtual machine deployment models for research reproducibility.
| System Metric / Parameter | OpenFold2 Baseline (Prior GPU Clusters) | ElMerFold Production Run (El Capitan) | ElMerFold Benchmark Peak (El Capitan) |
|---|---|---|---|
| Node Allocation | Distributed GPU clusters | 10,800 nodes | 11,200 nodes |
| Processor Hardware | Discrete CPUs + PCIe GPUs | 43,200 AMD Instinct MI300A APUs | 44,800 AMD Instinct MI300A APUs |
| Memory Subsystem | Split Host DDR + VRAM | Unified HBM3 (128 GB per APU) | Unified HBM3 (128 GB per APU) |
| Inference Throughput | ~147 structures / second | 2,400 structures / second | 6,718 structures / second |
| FP32 Computational Rate | Sub-exascale (under 50 PFLOP/s) | 604 PFLOP/s sustained | 1.81 Eflop/s sustained (1.99 Eflop/s peak) |
| Distillation Dataset | Under 1 million proteins | ~41 million monomer structures | ~41 million monomer structures |
| End-to-End Runtime | 6 to 8 days (projected) | ~11 hours | 8 to 11 hours |
Background and Context: The Self-Distillation Bottleneck
Modern biomolecular foundation models like AlphaFold and OpenFold rely heavily on ‘self-distillation’. Because experimentally determined structures in the Protein Data Bank (PDB) number fewer than 250,000 complexes, training frontier architectures requires generating millions of high-confidence synthetic structures from uncharacterized sequence databases such as MGnify. Generating this synthetic training data has historically formed a severe bottleneck, requiring three to ten times more compute than training the final neural network.
While DeepMind demonstrated multi-molecule complex prediction with AlphaFold3 in 2024, the model was released under closed proprietary licensing that restricted retraining, commercial deployment, and national security applications. Columbia University’s OpenFold consortium launched an initiative to build an open-source equivalent. ElMerFold directly resolves the distillation barrier for OpenFold3, generating the 41-million-structure synthetic library necessary to train the model openly.
The transition toward open structural AI highlights key governance questions regarding research access, safety evaluations, and benchmark reproducibility, as examined in CASRAI’s evaluation of open-weight AI models and scientific safety considerations. Disseminating multi-terabyte synthetic structural datasets also demands adherence to FAIR data principles, ensuring public outputs remain findable, accessible, interoperable, and reusable.
Stakeholder Reactions and Industry Perspectives
The achievement drew significant attention from leadership across supercomputing and biological research sectors. Nicholas Malaya, fellow in high-performance computing and sovereign AI at AMD, emphasized the hardware milestone: ‘ElMerFold showcases how the AMD-unified APU architecture accelerates the most demanding AI workloads at the largest scale ever achieved. Running on the world’s fastest supercomputer, this collaboration underscores what is possible when innovative hardware and leading-edge science come together to push the state of the art in AI.’
Nikoli Dryden, LLNL computer scientist and ElMerFold technical lead, highlighted the workflow transformation: ‘It is a more complex model, and there is a lot of complex orchestration around the model that you do not see in typical inference workflows. In some ways, it looks more like some of the scientific workflows that the lab does in other contexts. Many other applications are starting to look at this sort of large-scale inference workload, and being able to produce the OpenFold3 model is a centerpiece of a lot of planned initiatives and capabilities.’
OpenFold researchers noted that unblocking distillation democratizes structural biology. By pairing public data releases with open-source code, academic laboratories gain the ability to explore pathogen structures, study neglected diseases, and engineer enzymes without proprietary cloud restrictions.
What It Means for Researchers, Universities, and Research Offices
The demonstration of exascale AI distillation carries strategic implications for research universities and computing centers:
- Rebalancing Supercomputing Workloads: Academic supercomputing consortia have long prioritized PDE simulation and molecular dynamics. ElMerFold establishes that large-scale synthetic data distillation is now an essential HPC workload, requiring research computing directors to optimize queue policies and near-node storage for high-throughput inference pipelines.
- Democratizing Structural Biology Research: Very few academic departments possess massive GPU clusters. Open distribution of the 41-million-protein distillation dataset enables university laboratories to fine-tune specialized models or explore structural candidates on modest institutional servers without paying proprietary API charges.
- Research Security and Dual-Use Governance: Frontier biological foundation models intersect with federal biosecurity directives and export controls under America’s AI Action Plan. University research compliance officers and institutional biosafety committees (IBCs) must evaluate data governance and access controls surrounding synthetic macromolecular prediction libraries.
- Structural Data Stewardship: Curating petabyte-scale synthetic libraries requires clear institutional data stewardship. Research offices and libraries must systematically differentiate between empirical crystallographic coordinates and synthetic AI predictions, enforcing metadata standards as detailed in CASRAI’s analysis of scientific data versus structural metadata lifecycle management.
What Happens Next: Tuolumne Operations and OpenFold3 Training
With El Capitan transitioning to classified national security workloads for the NNSA’s Advanced Simulation and Computing (ASC) stockpile stewardship program, open-science biological workflows are transferring to Tuolumne. Rated at 288 petaflops, Tuolumne serves as El Capitan’s unclassified companion system, utilizing the same MI300A APU architecture at roughly one-tenth the scale.
The LLNL, Columbia, and AMD team will use Tuolumne to finalize the training of OpenFold3, targeting fully open model weights and code releases for the scientific community. The team also plans to extend the ElMerFold framework to model multimers and multi-protein complexes, scaling toward simulating upwards of one billion biomolecular interactions. The Gordon Bell Prize finalist presentation is scheduled for November 17, 2026, at SC26, where the ACM committee will evaluate the research alongside competing exascale achievements.
Sources
- Lawrence Livermore National Laboratory. (2026-11-15). ‘LLNL and partners launch record-breaking protein-folding workflow on world’s fastest supercomputer.’ https://www.llnl.gov/article/53581/llnl-partners-launch-record-breaking-protein-folding-workflow-worlds-fastest-supercomputer
- Lawrence Livermore National Laboratory Computing. (2026-11-15). ‘ElMerFold: Exascale Distillation Workflows for Protein Structure Prediction.’ https://computing.llnl.gov/projects/elmerfold
- SC26 Conference Program. (2026-11-15). ‘Exascale Distillation Workflows for Protein Structure Prediction on El Capitan (Session gb106).’ https://sc26.conference-program.com/presentation/?id=gb106&sess=sess218
- Advanced Micro Devices (AMD). (2026-11-15). ‘ElMerFold: Exascale AI for Protein Structure Prediction with El Capitan.’ https://www.amd.com/en/corporate/events/advancing-ai/sessions-catalog/elmerfold–exascale-ai-for-protein-structure-prediction-with-el-capitan.html
- Supercomputing News. (2026-11-15). ‘AI for Science Depends on Simulation Supercomputing: Gordon Bell Finalists Showcase Exascale Convergence.’ https://www.supercomputing.news/hpc/ai-for-science-depends-on-simulation-supercomputing







